Code-excited linear prediction

CELP represents speech as an excitation signal passed through a time-varying linear-prediction filter. Linear predictive coding estimates the short-term spectral envelope of speech, corresponding approximately to resonances of the vocal tract; the remaining signal is modeled as excitation associated with voiced, unvoiced, and transitional sounds.¹ Instead of transmitting every excitation sample, the encoder stores or generates a finite set of candidate vectors called a codebook. For each analysis frame, it selects an entry and a gain, then sends the codebook index, gain, and filter parameters. The decoder performs the same synthesis operation, so it needs only the compact parameters rather than the original waveform. This separation of spectral shape from excitation is the defining idea behind CELP.

The encoder chooses parameters by searching for the combination whose reconstructed speech most closely matches the input. This process, called analysis-by-synthesis, runs candidate excitation vectors through the synthesis filter and evaluates the resulting error with a perceptually weighted criterion.² A typical CELP design uses a fixed, or stochastic, codebook together with an adaptive codebook that represents periodicity from earlier excitation, making it effective for voiced speech and pitch repetition. The search is computationally demanding because many codebook entries, gains, and delay values may be tested. Practical implementations therefore exploit algebraic structure, subframe processing, fast correlation methods, and staged searches. The transmitted parameters are optimized for perceived speech quality, not merely for minimum sample-by-sample error.

CELP became a foundation for low-bit-rate telephony because it can preserve intelligibility and naturalness with far fewer bits than waveform coders. Standardized descendants include Algebraic Code-Excited Linear Prediction (ACELP), which replaces a large stored stochastic codebook with sparse pulse patterns, and various hybrid schemes that combine adaptive and fixed-codebook excitation.³ ITU-T G.729, for example, specifies an eight-kilobit-per-second speech codec based on conjugate-structure ACELP, while G.723.1 provides selectable low-rate modes for multimedia communication.⁴ CELP-family methods also appeared in mobile standards, digital enhanced cordless telephony, voice-over-IP equipment, and speech-storage systems. Their strengths are low delay, moderate complexity relative to later neural methods, and graceful operation at constrained bit rates.

CELP is not a single codec but a broad design family whose members differ in codebook construction, perceptual weighting, frame structure, and parameter quantization. The term can therefore describe both the original stochastic-codebook approach and related standardized variants such as ACELP. A subtle limitation is that the model is optimized for speech: music, applause, environmental noise, and simultaneous speakers may sound artificial or suffer from instability because they do not fit a single vocal-tract-plus-excitation model. Comfort-noise generation and voice-activity detection are commonly added around the core coder to avoid transmitting full-rate speech parameters during silence.⁵ CELP also influenced later wideband and transform-assisted codecs, even when their architectures no longer use a conventional fixed codebook.

Linear predictive coding

A method that predicts the current speech sample from preceding samples and represents the prediction error with a compact parameter set.

Codebook

A collection of candidate excitation vectors or pulse patterns addressed by transmitted indices.

Analysis-by-synthesis

An encoding strategy that evaluates candidate parameters after reconstructing the signal they would produce.

Adaptive codebook

A memory of earlier excitation used to model pitch periodicity and long-term speech correlation.

ACELP

Algebraic Code-Excited Linear Prediction, a CELP-family method using structured sparse excitation patterns.